Mixed-Script Speech Recognition Training via Selective Transliteration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems struggle with accurately transcribing and training on speech or written language that includes words from multiple languages or dialects, particularly due to difficulties in handling mixed scripts, leading to inconsistent transcription and reduced model learning opportunities.
Innovation Solution
A method involving transliteration of language examples into a single script to normalize training data, using finite state transducer networks and selective transliteration to maintain context and accuracy, with error rates adjusted to account for equivalent words in different scripts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If mixed script language examples are excluded from training data sets, then transcription consistency is improved, but the amount of training examples and model learning opportunities is reduced
Solution Approach 1:
The patent introduces transliteration as an intermediary process that converts words from secondary scripts into the primary script. This mediator enables mixed script examples to be included in training data while maintaining transcription consistency, as the transliterated form provides a standardized representation that the speech recognition model can learn from uniformly.
Solution Approach 2:
The patent changes the script parameter of out-of-script words by transliterating them into the primary script. This parameter transformation allows the same semantic content to be represented in a consistent script form, enabling both transcription consistency and retention of training examples to be achieved simultaneously.
2Quantity of substance
If mixed script language examples are included in training data sets without transliteration, then the amount of training examples is increased, but transcription consistency and model accuracy deteriorate
Solution Approach 1:
Transliteration serves as a preprocessing intermediary that standardizes the script representation of training examples before they are fed to the speech recognition model. This ensures that while diverse script examples are included, they are all converted to a consistent primary script form, maintaining transcription consistency.
Solution Approach 2:
The patent applies transliteration as a preliminary action before training the speech recognition model. By pre-converting all words to the primary script, the training data is prepared in a consistent format that enables the model to learn uniform transcription patterns while still benefiting from the increased diversity and quantity of training examples.
3Manufacturing precision
If transliteration is applied to all words, then transcription consistency is improved, but proper names and certain contextual information may be lost
Solution Approach 1:
The patent applies transliteration selectively rather than uniformly to all words. By identifying and preserving proper names and other contextual information that should maintain their original script form, while transliterating only the necessary out-of-script words, the system maintains local quality differences that preserve important information while achieving transcription consistency where needed.
Solution Approach 2:
The patent applies parameter changes (transliteration) selectively based on word type and context. Rather than changing the script parameter for all words, it only modifies the parameter for out-of-script words that require normalization, while leaving proper names and contextually important words in their original script form to preserve information.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, systems, and apparatus, including computer programs stored on a computer-readable storage medium, for transliteration for speech recognition training and scoring. In some implementations, language examples are accessed, some of which include words in a first script and words in one or more other scripts. At least portions of some of the language examples are transliterated to the first script to generate a training data set. A language model is generated based on occurrences of the different sequences of words in the training data set in the first script. The language model is used to perform speech recognition for an utterance.